Since your fastCompletionStage implementation was just a pass through, how do you expect to implement the other methods on the completion stage and still not take a performance hit on the per element call since the fastCompletionStage did not actually implement any of the other methods?
On Sun, Mar 25, 2018 at 10:20 AM Romain Manni-Bucau <rmannibucau@xxxxxxxxx> wrote:@Lukasz: just a small precision on the bench I shared earlier: the overhead of CompletionStage (implemented with a "fast" flavor) is of < 7% if you ignore the usage of lambda (pass a function instance and not using lambda ref - not sure why the JVM doesn't handles it directly but since a JVM upgrade from the u40 to u144 made a 75% boost thanks to lambda+gc optims, I don't worry much of that part). Here are the raw results I get (Sharing beam one too since I used another computer):Comparison.beam thrpt 5 184033706,109 ± 31943851,553 ops/sComparison.
fastCompletionStageWithoutLamb da thrpt 5 171628984,800 ± 2063217,863 ops/sI insist on the good fit of CompletionStage (or any reactive compatible API closer to java 9 maybe) but I had to migrate from a synchronous code to an async one on friday and the migration was not technically hard and brought a lot of benefit since now it can work in any environment (synchronous using toCompletionFuture().get() or asynchronous like akka actors bridging scala future and CompletionStage). For a portable API (I'm not speaking of the beam - language - portable API which is on top of runner from a design point of view) but of the API any runner must integrate with. Integrated with IO (which is the only part giving sense to any pipeline when you think about it) you can scala way more reliable and efficiently optimizing your resources so it would be an awesome fit for a solution like beam IMHO.2018-03-15 18:45 GMT+01:00 Jean-Baptiste Onofré <jb@xxxxxxxxxxxx>:By the way, you can take a look on JdbcIO which does a reshuffle transform to avoid the "fusion" issue.RegardsJBLe 15 mars 2018, à 10:44, Raghu Angadi <rangadi@xxxxxxxxxx> a écrit:In streaming, a simple way is to add a reshuffle to increase parallelism. When you are external-call bound, extra cost of reshuffle is negligible. e.g. https://stackoverflow.com/ questions/46116443/dataflow- streaming-job-not-scaleing- past-1-worker
Note that by default Dataflow workers use a couple of hundred threads as required. This can be increased with a pipeline option if you prefer. I am not sure of other runners.
On Thu, Mar 15, 2018 at 8:25 AM Falcon Taylor-Carter < falcon@xxxxxxxxxxxxxxxxxx> wrote:
Thanks for checking up (I'm working with Josh on this problem). It seems there isn't a built-in process for this kind of use case currently, and that the best process right now is to handle our own bundling and threading in the DoFn. If you had any other suggestions, or anything to keep in mind in doing this, let us know!
On Tue, Mar 13, 2018 at 4:52 PM, Pablo Estrada <pabloem@xxxxxxxxxx> wrote:
I'd just like to close the loop. Josh, did you get an answer/guidance on how to proceed with your pipeline?--Or maybe we'll need a new thread to figure that out : )Best-P.
On Fri, Mar 9, 2018 at 1:39 PM Josh Ferge < josh.ferge@xxxxxxxxxxxxxxxxxx> wrote:
Our team has a pipeline that make external network calls. These pipelines are currently super slow, and the hypothesis is that they are slow because we are not threading for our network calls. The github issue below provides some discussion around this:
In beam 1.0, there was IntraBundleParallelization, which helped with this. However, this was removed because it didn't comply with a few BEAM paradigms.
Questions going forward:
What is advised for jobs that make blocking network calls? It seems bundling the elements into groups of size X prior to passing to the DoFn, and managing the threading within the function might work. thoughts?Are these types of jobs even suitable for beam?Are there any plans to develop features that help with this?
Got feedback? go/pabloem-feedback